WSP-Bench dev instances (seed 42): A11-3

README.md

@@ -32,6 +32,7 @@ return outputs, L3 outputs plus activations and attention maps, L4 interventions
 | A10-1 | Policy-Failure | 1 |
 | A10-2 | Injected-Defect | 40 |
 | A10-3 | LM-Bug | 1 |
+| A11-3 | VLA-Finetune-ID | 1 |
 | A2-1 | DDPM-Shapes | 25 |
 | A2-2 | Decoder-Factors | 28 |
 | A2-3 | CharGen-Format | 1 |

manifest.json

@@ -3340,6 +3340,33 @@
    "gpu": "NVIDIA A100-SXM4-80GB"
   }
  },
+ {
+  "task": "A11-3",
+  "name": "VLA-Finetune-ID",
+  "D": 3,
+  "Z": 11,
+  "seed": "42.0",
+  "iid": "13f207c1",
+  "n_params": 450046176,
+  "s1t_format": null,
+  "path": "tasks/A11-3/D3/Z11/13f207c1",
+  "files": {
+   "base.safetensors": "ed5df079ac044d34dd44b5f89ea4db1f98aaf7cf3463e93da11a0a2bc7a5c131",
+   "materials/practice.json": "db5b5bfd5fbd0b3d247a7d3b31fb4cb030c93fe1d7d1139ce8817ce6d0643b8d",
+   "materials/practice.safetensors": "e0c73680564ce9a4c0c3b4e4ffeac3f8401ada5b7eacbcc8f042867a02c20a00",
+   "question.md": "58460d9a188d0386271dc328d1facba7290f3b994d8c8bfcee265f6a6f0006fe",
+   "subject.md": "a540162423293d109228a598927e28301276b77015d10237ab7e262c3a278ee8",
+   "tokenizer/tokenizer.json": "0db9d5da9b0313acf440ff3cd4c30402c7bd16bbbddbde520dc058c813ea70a1",
+   "tuned.safetensors": "9c88629e19d593fcc05d8fcb46671dcf04f2f27557c374cd67f2ad52300c306e"
+  },
+  "built_with": {
+   "torch": "2.6.0+cu124",
+   "numpy": "2.4.6",
+   "python": "3.11.18",
+   "cuda": "12.4",
+   "gpu": "NVIDIA A100-SXM4-80GB"
+  }
+ },
  {
   "task": "A2-1",
   "name": "DDPM-Shapes",

tasks/A11-3/D3/Z11/13f207c1/base.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ed5df079ac044d34dd44b5f89ea4db1f98aaf7cf3463e93da11a0a2bc7a5c131
+size 900148584

tasks/A11-3/D3/Z11/13f207c1/materials/practice.json

@@ -0,0 +1,36 @@
+{
+ "frames": 20,
+ "cameras": [
+  "camera1"
+ ],
+ "instructions": [
+  "Push the T-shaped block onto the T-shaped target.",
+  "Push the T-shaped block onto the T-shaped target.",
+  "Push the T-shaped block onto the T-shaped target.",
+  "Push the T-shaped block onto the T-shaped target.",
+  "Push the T-shaped block onto the T-shaped target.",
+  "Push the T-shaped block onto the T-shaped target.",
+  "Push the T-shaped block onto the T-shaped target.",
+  "Push the T-shaped block onto the T-shaped target.",
+  "Push the T-shaped block onto the T-shaped target.",
+  "Push the T-shaped block onto the T-shaped target.",
+  "Push the T-shaped block onto the T-shaped target.",
+  "Push the T-shaped block onto the T-shaped target.",
+  "Push the T-shaped block onto the T-shaped target.",
+  "Push the T-shaped block onto the T-shaped target.",
+  "Push the T-shaped block onto the T-shaped target.",
+  "Push the T-shaped block onto the T-shaped target.",
+  "Push the T-shaped block onto the T-shaped target.",
+  "Push the T-shaped block onto the T-shaped target.",
+  "Push the T-shaped block onto the T-shaped target.",
+  "Push the T-shaped block onto the T-shaped target."
+ ],
+ "state": "normalised robot state, (x - mean) / std of its dataset",
+ "actions": "the recorded 50-step action chunk starting at the frame, normalised the same way",
+ "images": "uint8 RGB, (frames, height, width, 3)",
+ "about": "a simulated 2-D pushing task (a round pusher pushes a T-shaped block onto a target), one top-down camera; state and action are the 2-D pusher position",
+ "slots": {
+  "image": "camera1"
+ },
+ "note": "frames of a robot dataset that is none of the five candidates"
+}

tasks/A11-3/D3/Z11/13f207c1/materials/practice.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e0c73680564ce9a4c0c3b4e4ffeac3f8401ada5b7eacbcc8f042867a02c20a00
+size 561344

tasks/A11-3/D3/Z11/13f207c1/question.md

@@ -0,0 +1,10 @@
+YOUR TASK. tuned.safetensors was made from base.safetensors by fine-tuning on the demonstrations of exactly one of the five public robot datasets below. Which one?
+CANDIDATES.
+A. Robot: a real SO-101 arm (the joint layout of SO-100, a different build). Task: put a small pink Lego brick into a transparent box; instruction: "pink lego brick into the transparent box". Cameras: top view -> slot 1 (640x480), side view -> slot 3 (640x480), 30 frames per second. 50 demonstrations, 11,939 frames. State: 6 numbers (shoulder_pan.pos, shoulder_lift.pos, elbow_flex.pos, wrist_flex.pos, wrist_roll.pos, gripper.pos); action: 6 numbers (shoulder_pan.pos, shoulder_lift.pos, elbow_flex.pos, wrist_flex.pos, wrist_roll.pos, gripper.pos).
+B. Robot: a simulated Franka Panda arm with a parallel gripper (LIBERO benchmark scenes). Task: 40 kitchen / tabletop manipulation tasks from four task suites (spatial layouts, different objects, different goals, long two-step tasks); 40 instructions, for example "put the white mug on the left plate and put the yellow and white mug on the right plate"; "put the cream cheese in the bowl"; "pick up the salad dressing and place it in the basket"; "pick up the black bowl on the wooden cabinet and place it on the plate". Cameras: third-person view -> slot 1 (256x256), wrist view -> slot 2 (256x256), 10 frames per second. 1693 demonstrations, 273,465 frames. State: 8 numbers; action: 7 numbers.
+C. Robot: a real SO-100 arm (5 revolute joints and a gripper). Task: pick up the red cube and put it on top of the blue cube; the start positions of both cubes vary; instruction: "Put the red cube on top of the blue cube.". Cameras: top view -> slot 1 (640x480), wrist view -> slot 2 (640x480), 30 frames per second. 56 demonstrations, 22,956 frames. State: 6 numbers (main_shoulder_pan, main_shoulder_lift, main_elbow_flex, main_wrist_flex, main_wrist_roll, main_gripper); action: 6 numbers (main_shoulder_pan, main_shoulder_lift, main_elbow_flex, main_wrist_flex, main_wrist_roll, main_gripper).
+D. Robot: a real SO-100 arm (5 revolute joints and a gripper). Task: pick up a cube and put it in a small box; the box stays at one place, the cube starts at one of five positions; instruction: "Pick up the cube and place it in the box.". Cameras: top view -> slot 1 (640x480), wrist view -> slot 2 (640x480), 30 frames per second. 50 demonstrations, 19,631 frames. State: 6 numbers (main_shoulder_pan, main_shoulder_lift, main_elbow_flex, main_wrist_flex, main_wrist_roll, main_gripper); action: 6 numbers (main_shoulder_pan, main_shoulder_lift, main_elbow_flex, main_wrist_flex, main_wrist_roll, main_gripper).
+E. Robot: a real SO-100 arm (5 revolute joints and a gripper). Task: a longer task: put the red cube in the right box and the blue cube in the left box; the cubes start at five positions, with their colours swapped between demonstrations; the boxes stay at fixed places; instruction: "Put the red cube in the right box and the blue cube in the left box.". Cameras: top view -> slot 1 (640x480), wrist view -> slot 2 (640x480), 30 frames per second. 52 demonstrations, 35,713 frames. State: 6 numbers (main_shoulder_pan, main_shoulder_lift, main_elbow_flex, main_wrist_flex, main_wrist_roll, main_gripper); action: 6 numbers (main_shoulder_pan, main_shoulder_lift, main_elbow_flex, main_wrist_flex, main_wrist_roll, main_gripper).
+FINE-TUNING (the same whatever the dataset). 6000 steps of 32 frames drawn in a random order from the dataset's training demonstrations (20% of its demonstrations were held out and not used); each frame with its camera images (in the slots given above), its instruction, its state and the next 50 recorded actions, state and actions normalised with that dataset's own mean and standard deviation. The flow-matching loss (velocity error at a random flow time, averaged over the 50 x 32 padded chunk); AdamW (learning rate 0.0001, betas 0.9 / 0.95, weight decay 1e-10), 100 warm-up steps then cosine decay to 2.5e-06, gradient norm clipped at 10. Only the action expert (expert.*), state_proj and the action / time projections were trained; vision.*, connector.* and vlm.* were frozen.
+ANSWER. Reply with JSON: {"label": "<one of A-E>", "probabilities": {"A": p, "B": p, "C": p, "D": p, "E": p}} (the probabilities are normalised to sum 1). First explain your evidence.
+SCORE. 1 if your label is the dataset that was used, else 0. The Brier score of your probabilities is reported as well, so let them say how sure you are.

tasks/A11-3/D3/Z11/13f207c1/subject.md

@@ -0,0 +1,12 @@
+THE NETWORKS
+Two vision-language-action robot policies of one architecture, 450,046,176 parameters each, in bfloat16 with the same tensor names and shapes: base.safetensors (a pretrained policy) and tuned.safetensors (the same policy after fine-tuning). Both files use the same order of heads and neurons, so row / column i of a tensor is the same unit in both. Given camera images, an English instruction and the robot state, the policy outputs a chunk of 50 future actions.
+ARCHITECTURE (tensor names follow it; a linear layer's weight has shape (out, in)).
+- vision.*: a ViT image encoder (12 layers, width 768, 12 heads, 16x16 patches of a 512x512 image). connector.*: merges each 4x4 block of patch features and projects it to width 960: 64 tokens per camera image.
+- vlm.*: a decoder-only language model truncated to 16 layers (width 960, 15 query heads and 5 key/value heads of size 64, rotary positions, RMSNorm, gated SiLU MLPs of width 2560); vlm.embed_tokens embeds the instruction; vlm.lm_head is not used by the policy.
+- state_proj: the robot state (zero-padded to 32 numbers) -> one token of width 960. The prefix is the image tokens of each camera, then the instruction tokens, then the state token.
+- expert.*: the action expert, 16 layers of width 720 (15 query heads, 5 key/value heads of size 64, gated SiLU MLPs of width 2048), run on 50 action tokens. In layers [0, 2, 4, 6, 8, 10, 12, 14] the prefix and the action tokens attend together (each stream with its own projections; the prefix never attends to the action tokens). In layers [1, 3, 5, 7, 9, 11, 13, 15] the action tokens attend only to the language model's keys and values of the same layer, mapped by the expert's own k_proj / v_proj (320 -> 320).
+- action_in_proj, action_time_mlp_in, action_time_mlp_out: the noisy chunk (50 x 32) and a sinusoidal embedding of the flow time t (periods 0.004 to 4.0) -> the action tokens; action_out_proj: the expert's output -> a velocity (50 x 32).
+SAMPLING. Flow matching: x(1) is standard normal noise (50 x 32); 10 Euler steps x(t - 0.1) = x(t) - 0.1 v(x(t), t) from t = 1 to t = 0; the first action_dim numbers of each of the 50 rows of x(0) are the normalised actions of the next 50 steps. Training target: v = noise - action at x(t) = t noise + (1 - t) action.
+INPUTS. Each camera image (RGB in [0, 1]) is resized with its aspect ratio kept to fit 512x512, zero-padded and mapped to [-1, 1]. Cameras fill up to three slots in a fixed order (slot 1: top / third-person view, slot 2: wrist view, slot 3: side view); a slot without a camera is left out. The instruction, with a newline appended, is tokenised with tokenizer/tokenizer.json and right-padded / cut to 48 tokens (padding masked). State and action are normalised per dataset, (x - mean) / std, and zero-padded to 32 / 32 numbers.
+The order of the MLP neurons and of the query heads inside each key/value group of every vlm.* and expert.* layer is arbitrary.
+FILES. base.safetensors, tuned.safetensors, tokenizer/tokenizer.json (the instruction tokenizer). materials/practice.safetensors and materials/practice.json: 20 practice frames of another robot dataset (a simulated 2-D pushing task (a round pusher pushes a T-shaped block onto a target), one top-down camera; state and action are the 2-D pusher position) that is none of the candidates in the question: camera images (uint8, frames x height x width x 3, keyed by slot), the normalised state, the recorded 50-step action chunk (normalised), and the instruction. Where you have a forward tool, these are the inputs it runs on (you may change their instructions or other inputs as the tool allows).

tasks/A11-3/D3/Z11/13f207c1/tuned.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9c88629e19d593fcc05d8fcb46671dcf04f2f27557c374cd67f2ad52300c306e
+size 900148584